[PATCH 3/5] nvme-rdma: Make sure to shutdown the controller if we can

[sagi@grimberg.me (Sagi Grimberg)]

>>>> -	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
>>>> +	if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
>>>>  		nvme_shutdown_ctrl(&ctrl->ctrl);
>>>>
>>>>  	blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
>>>
>>> Maybe the right way to handle this is to unconditionally call
>>> nvme_shutdown_ctrl and make sure we return an early error
>>> on the register write?
>>
>> As I wrote on patch 2/5 reply, I'd like to avoid depending on
>> queue_rq to fail early peeking at the ctrl state. This dependency
>> can grow in the future and I think we should at least try not to
>> go there...
>
> I don't want ->queue_rq to peek at controller state.  What I had
> in mind was copying the PCIe behavior of failing early in the
> timeout handler if we are in the reset handler, which will
> mean blk_execute_rq will return ASAP with an error.

Do we want the unnecessary ADMIN_TIMEOUT for sending a shutdown
on a queue we know is not connected? I'm worried it will have scale
issues...